Add Ernie 4.5 OpenVINO support (#1366)

* rebase to non-remote model

* update min transformers lib for ernie 4.5

* add export test

* switch to OVDecoderModelPatcher

Author: openvino-dev-samples

docs/source/openvino/models.mdx

@@ -47,6 +47,7 @@ Here is the list of the supported architectures :
 - Deepseek_v2
 - Deepseek_v3
 - DistilBert
+- Ernie4.5
 - Electra
 - Encoder Decoder
 - ESM

optimum/exporters/openvino/model_configs.py

@@ -4512,3 +4512,17 @@ def generate_dummy_inputs(self, framework: str = "pt", **kwargs):
                 )
 
         return dummy_inputs
+
+
+@register_in_tasks_manager("ernie4_5", *["text-generation", "text-generation-with-past"], library_name="transformers")
+class ErnieOpenVINOConfig(TextDecoderWithPositionIdsOnnxConfig):
+    MIN_TRANSFORMERS_VERSION = "4.54.0"
+
+    DUMMY_INPUT_GENERATOR_CLASSES = (DummyTextInputGenerator, GemmaDummyPastKeyValuesGenerator)
+    DUMMY_PKV_GENERATOR_CLASS = GemmaDummyPastKeyValuesGenerator
+    NORMALIZED_CONFIG_CLASS = NormalizedTextConfig
+
+    def patch_model_for_export(
+        self, model: Union["PreTrainedModel", "TFPreTrainedModel"], model_kwargs: Optional[Dict[str, Any]] = None
+    ) -> "ModelPatcher":
+        return OVDecoderModelPatcher(self, model, model_kwargs=model_kwargs)

tests/openvino/test_export.py

@@ -98,6 +98,9 @@ class ExportModelTest(unittest.TestCase):
             {"stable-diffusion-3": OVStableDiffusion3Pipeline, "flux": OVFluxPipeline, "ltx-video": OVLTXPipeline}
         )
 
+    if is_transformers_version(">=", "4.54"):
+        SUPPORTED_ARCHITECTURES.update({"ernie4_5": OVModelForCausalLM})
+
     GENERATIVE_MODELS = ("pix2struct", "t5", "bart", "gpt2", "whisper", "llava", "speecht5")
 
     def _openvino_export(

tests/openvino/test_exporters_cli.py

@@ -114,6 +114,14 @@ class OVCLIExportTestCase(unittest.TestCase):
                 ("text-to-video", "ltx-video"),
             ]
         )
+
+    if is_transformers_version(">=", "4.54"):
+        SUPPORTED_ARCHITECTURES.extend(
+            [
+                ("text-generation-with-past", "ernie4_5"),
+            ]
+        )
+
     EXPECTED_NUMBER_OF_TOKENIZER_MODELS = {
         "gpt2": 2 if is_tokenizers_version("<", "0.20") or is_openvino_version(">=", "2024.5") else 0,
         "t5": 0 if is_openvino_version("<", "2025.1") else 2,  # 2025.1 brings support for unigram tokenizers
@@ -137,6 +145,7 @@ class OVCLIExportTestCase(unittest.TestCase):
         "clip": 2 if is_tokenizers_version("<", "0.20.0") or is_openvino_version(">=", "2024.5") else 0,
         "mamba": 2,
         "falcon-mamba": 2,
+        "ernie4_5": 2,
     }
 
     TOKENIZER_CHAT_TEMPLATE_TESTS_MODELS = {

tests/openvino/test_modeling.py

@@ -1183,6 +1183,9 @@ class OVModelForCausalLMIntegrationTest(unittest.TestCase):
     if is_transformers_version(">=", "4.51.3"):
         SUPPORTED_ARCHITECTURES += ("glm4",)
 
+    if is_transformers_version(">=", "4.54.0"):
+        SUPPORTED_ARCHITECTURES += ("ernie4_5",)
+
     GENERATION_LENGTH = 100
     REMOTE_CODE_MODELS = (
         "chatglm",
@@ -1270,6 +1273,7 @@ class OVModelForCausalLMIntegrationTest(unittest.TestCase):
         "qwen3_moe": 2,
         "mamba": 0,
         "falcon-mamba": 0,
+        "ernie4_5": 2,
     }
 
     # TODO: remove gptq/awq from here

tests/openvino/utils_tests.py

@@ -63,6 +63,7 @@
     "donut": "fxmarty/tiny-doc-qa-vision-encoder-decoder",
     "donut-swin": "hf-internal-testing/tiny-random-DonutSwinModel",
     "detr": "hf-internal-testing/tiny-random-DetrModel",
+    "ernie4_5": "optimum-internal-testing/tiny-random-Ernie4_5ForCausalLM",
     "electra": "hf-internal-testing/tiny-random-electra",
     "esm": "hf-internal-testing/tiny-random-EsmModel",
     "exaone": "katuni4ka/tiny-random-exaone",